Skip to content

AI

NYT says OpenAI claimed it couldn't search its own data. Then it did.

The New York Times and Daily News allege OpenAI pleaded "we can't search that" for two years while quietly running its own searches. OpenAI's side of the story is a privacy-and-burden argument.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. NYT and Daily News allege OpenAI misrepresented whether it can search its training data.
  2. A deposition allegedly shows OpenAI engineers had already run internal copyright searches.
  3. OpenAI's stated defense: searching chat logs is burdensome and risks user privacy.

What happened

The lawsuit is roughly two years old. The outlets accuse OpenAI of training its generative models on Times content without permission and of reproducing that journalism in what ChatGPT tells users.

To prove it, the plaintiffs wanted access to two things: the training dataset, and the huge pile of ChatGPT conversations. The goal is to see whether their articles are in there and how often ChatGPT draws on or repeats them.

According to the filings, OpenAI had argued it could not search the data it trained on. It also argued that searching or handing over the chat logs would be technically burdensome. On top of that, it raised user privacy, since the logs would have to be retrieved, processed and de-identified first.

Then came an April court-ordered deposition. The plaintiffs say OpenAI data privacy engineer Vinnie Monaco let slip that the company had already run internal searches and evaluations of its training corpus, hunting for copyrighted journalism.

If that holds up, "we can't" starts to look a lot like "we did, just not for you."

Why it matters

Discovery is where AI copyright cases are won or lost. Whether a model absorbed protected work is a question of evidence, and the evidence sits on the defendant's servers.

If a company can search its data for its own purposes but tells a court it cannot search for the plaintiff's, judges tend to notice. This is still an allegation from one side. But it is aimed squarely at OpenAI's credibility on a core factual point, in one of the most closely watched copyright cases in AI.

It also puts the privacy argument under a brighter light. Protecting user conversations is a legitimate concern. It is also a very convenient shield, and courts will weigh whether it is doing real work here.

For everyone else training models on scraped content, the lesson is plain: your data may eventually be subpoenaed, and "it's too big to look at" is a shaky defense.

What to do with it

  • Watch the court, not the press release.The next filings and any judge's response will show whether the allegation sticks.
  • Know your own data trail.If you build or buy AI, ask vendors what they can actually search and audit in their training data.
  • Treat chat logs as discoverable.Anything your team pastes into a chatbot may one day be retrieved, processed and read by strangers.
  • Publishers, document everything.If your work is in a model, evidence of reproduction in outputs is what these cases run on.

Ciao, sapiens. When the machine says it cannot look, check who else already looked.